Action Gate
Check an agent’s exact action against your approved policy before it runs.
Your agent sends the tool call it intends to make. Action Gate answers yes or no against the policy version you approved, with a one-line reason. Anything ambiguous, missing or out of scope is a no.
You send
- — Once: your policy, which you approve after reviewing its explicit restatement.
- — Per action: the exact tool call — name, arguments, resources and relevant context.
You get
- — yes or no, with a short reason and the policy section it rests on.
- — A yes bound to that exact payload and policy version. Change the call and you need a new decision.
- — no — with the reason — whenever the policy, the facts or the interpretation are unsettled.
Illustrative example
Written by hand to show the shape of a result. Not a live or measured output.
You send
- Policy v4
- Support agents may issue refunds of up to $100 to the original payment method. (approved)
- Call
- issue_refund(amount: 80, currency: "USD", method: "original")
You get
- Decision
- yes — within the $100 limit, to the original payment method (v4 §1)
You send
- Policy v4
- Support agents may issue refunds of up to $100 to the original payment method. (approved)
- Call
- issue_refund(amount: 80, currency: "USD", method: "store_credit")
You get
- Decision
- no — refunds must go to the original payment method (v4 §1)
You send
- Policy v4
- Support agents may issue refunds of up to $100 to the original payment method. (approved)
- Call
- issue_refund(amount: 95, currency: "EUR", method: "original")
You get
- Decision
- no — the limit is stated in USD; the approved policy does not cover currency conversion
What it will not do
- — Return an implicit yes. Silence, timeouts and ambiguity are no.
- — Reuse a decision for a different payload or policy version.
- — Enforce anything on its own: your agent or proxy must act on the decision before executing.
- — Claim coverage it does not have. General temporal, numerical and legal policy reasoning is out of scope at first.
How you will connect
- HTTPS API.
- Submit a task, follow a durable job, fetch the result.
- MCP.
- One task-specific tool, gate_action, for agents and assistants.
- WebSocket.
- Progress and the final result for long tasks, resumable after a disconnect.
Results are readable statements, decisions and citations. Internal representations stay private.
Before launch
We are qualifying Action Gate with invited partners on their own material. Before general availability we will publish the workloads tested, error and abstention rates, and known limits. Until then, treat this page as a description of work in progress.
Bring a task to the pilot